OncoFirm™ / Research / Precision diagnostics
Pillar guide · Precision diagnostics
AI-Driven Biomarkers: How Algorithms Are Shaping the Future of Precision Diagnostics
Artificial intelligence is changing what a biomarker can be. Instead of reading one value against one cutoff, algorithms now combine several measurements, images or molecular patterns into a single, calibrated result. This guide explains what AI-driven biomarkers are, which algorithm-based tests are already in clinical use, how they earn trust, how the FDA oversees them, and why precise, quantitative measurement is the foundation everything else rests on.
At a glance
Topic
AI-driven biomarkers in precision diagnostics
In clinical use
Algorithm-based tests since 2009 (OVA1); AI pathology prognosis since 2025 (ArteraAI Prostate)
FDA landscape
More than 1,600 AI-enabled devices authorized as of September 2026
Core principle
An algorithm is only as reliable as the measurements it receives
OncoFirm™ status
Platform in development; not FDA cleared or approved
On this page
Key points
AI-driven biomarkers in brief
1,600+AI-enabled medical devices on the FDA's authorized list as of September 2026
2009first FDA clearance of a multi-protein algorithm test for ovarian cancer risk (OVA1)
3 testsof evidence every biomarker must pass: analytical validity, clinical validity, clinical utility
Figures are from FDA and company sources listed in the references.
Definition
What are AI-driven biomarkers?
An AI-driven biomarker is a measurable indicator of a biological state whose interpretation depends on an algorithm. The algorithm may combine several laboratory values with clinical details such as age into one risk score, or extract a pattern from complex data such as a digitized tissue slide, an imaging study or the fragment profile of cell-free DNA.
The underlying measurements are often familiar. Proteins such as CEA, CA-125, PSA and AFP have been measured for decades. What changes is the interpretation: a trained model weighs several signals together, can account for factors such as age or menopausal status, and can report how confident it is. For the basics, see what are cancer biomarkers?.
Types of biomarkers
Six jobs a biomarker can do
The FDA-NIH BEST glossary groups biomarkers by what they are used for. AI can support any of these roles, but the evidence needed differs: a screening or diagnostic claim must be tested in the people who will actually be screened, while a monitoring claim depends on precision across repeated measurements.
Diagnostic
Detects or confirms a disease, or identifies a subtype.
Monitoring
Measured repeatedly to track disease status or response over time.
Prognostic
Estimates the likely course of disease, such as recurrence risk.
Predictive
Identifies who is more likely to benefit from a specific treatment.
Susceptibility / risk
Indicates the chance of developing a disease in people without it.
Safety
Detects possible toxicity or adverse effects of an exposure or treatment.
How it works
From sample to trusted result
Every stage matters. A model can only be as good as the measurements it receives, and its performance must be checked continuously once it is in use.
- Collect. A blood, serum or plasma sample is taken under defined conditions, because handling and timing affect many biomarkers.
- Measure. Each biomarker is quantified in standard units, with calibration traceable to reference materials so results compare across sites and over time.
- Check. Built-in controls and quality flags catch invalid runs before any interpretation happens.
- Model. A locked, validated algorithm combines the values, and often clinical details, into a score with a stated level of confidence.
- Decide. A clinician uses the result alongside history, examination and other tests. The algorithm informs the decision; it does not make it.
Algorithms already in use
AI-driven biomarkers are not only a future idea
Combining several biomarkers with an algorithm has been part of cancer diagnostics for more than fifteen years. The newest tools add machine learning and new data types, but the principle is the same: several imperfect signals, weighed together, can be more informative than any one alone.
| Test | Inputs | What the algorithm does | U.S. status |
|---|---|---|---|
| OVA1 | Five serum proteins, including CA-125 | Assesses the likelihood of ovarian malignancy in a woman with a pelvic mass | FDA cleared, 2009 |
| ROMA | HE4, CA-125, menopausal status | Classifies an adnexal mass as high or low likelihood of malignancy to guide referral | FDA cleared, 2011 |
| GALAD score | Gender, age, AFP-L3, AFP, DCP | Supports earlier detection of hepatocellular carcinoma in at-risk patients | Breakthrough Device designation, 2020 |
| ArteraAI Prostate | Digitized biopsy images plus clinical data | Estimates long-term outcomes in non-metastatic prostate cancer | FDA De Novo, 2025 |
| CancerSEEK (research) | Cell-free DNA mutations plus eight proteins | Detects and localizes several cancer types from one blood draw | Research study, 2018 |
Notice how many of these tools rely on protein measurements. That is why quantitative protein testing remains central to precision diagnostics, including the multi-biomarker blood tests now being explored for triage and the cell-free DNA methods discussed in our ccfDNA research explainer.
Earning trust
What makes an AI-driven biomarker trustworthy
A model that looks accurate during development can still fail in practice. The most common reasons are well understood, and each has a known safeguard.
- Overfitting. Models trained on many variables and few patients can memorize noise. Safeguard: test on independent data that played no part in building the model.
- Spectrum bias. A test validated only against healthy volunteers can look better than it is. Safeguard: include people with look-alike conditions, such as inflammatory or benign disease.
- Measurement variation. Differences between labs, lots or devices shift inputs and therefore scores. Safeguard: calibration traceable to reference standards and ongoing quality control.
- Population differences. Performance can vary by age, sex, ancestry and setting. Safeguard: diverse validation cohorts and reporting of results by subgroup.
- Drift. Clinical practice, assays and populations change over time. Safeguard: performance monitoring after release and controlled, documented updates.
- Opacity. Clinicians need to know when to rely on a result. Safeguard: clear intended use, stated confidence and quality flags with every result.
These safeguards apply to any test, not only AI. They shape how OncoFirm™ approaches AI-assisted diagnostics and its evidence hub.
Regulation
How the FDA oversees AI-driven biomarkers
In the United States, software that interprets patient data to inform diagnosis is generally regulated as a medical device, either on its own or as part of a test system. The FDA’s public list of AI-enabled devices passed 1,600 authorizations by September 2026, with most in medical imaging so far. Algorithm-based laboratory tests such as OVA1 and ROMA reached the market through the same device pathways.
Adaptive software raises a particular question: what happens when a model is updated? In December 2024 the FDA finalized guidance on Predetermined Change Control Plans (PCCPs); the current version is dated August 2025. A PCCP lets a manufacturer describe planned model changes, and how each will be validated, in its original submission, so that pre-agreed updates can be made without a new submission each time.
For developers this means planning the full life cycle from the start: intended use, training and test data, performance targets, monitoring and change control. For health systems it means asking for the same evidence they would expect from any diagnostic. See what healthcare systems should know about point-of-care testing.
OncoFirm™ roadmap
Building the foundation for AI-driven biomarkers at the point of care
OncoFirm™’s view is simple: precision diagnostics start with precise measurement. Our platform is designed to deliver quantitative, traceable protein results in about 20 minutes outside the central laboratory, so that multi-marker algorithms can one day be used where patients are first seen.
Quantitative inputs first
A self-calibrating fluorescent strip designed to give numbers, not just lines: CEA over a designed 1–100 ng/mL range and PSA over 0.5–50 ng/mL, with planned traceability to WHO international standards. Platform
The reader as a data source
A handheld digital reader that captures each result with quality flags and a time stamp, creating structured data at the point of care. Digital readers
Locked, validated models
Interpretation software planned as locked versions with documented validation, and a change control plan for any future update.
Specificity by design
Validation plans that include benign and inflammatory conditions, not only healthy controls, across diverse populations. Clinical collaborations
Explore the full biomarker landscape in the OncoFirm™ Cancer Biomarker Atlas 2026 and how one reader and one strip design can support a growing menu of tests.
OncoFirm™ assays, reader and software are in development, have not been cleared or approved by the FDA, and are for research use only. Designed ranges and the 20-minute result time are development targets.
FAQ
Frequently asked questions
What is an AI-driven biomarker?
An AI-driven biomarker is a measurable indicator of health or disease whose interpretation depends on an algorithm. The algorithm may combine several lab values with clinical details into one score, or extract a pattern from data such as a digitized biopsy slide or a fragment profile of cell-free DNA.
How is an AI-driven biomarker different from a traditional one?
A traditional biomarker is usually read against a single cutoff, such as a PSA value. An AI-driven biomarker combines multiple inputs and weighs them with a trained model, which can capture patterns no single marker shows. It also brings new risks, including overfitting, hidden bias and performance drift, so it needs extra validation.
Are there FDA-authorized tests that use algorithms to combine biomarkers?
Yes. Examples include OVA1, cleared in 2009, which combines five serum proteins to assess the likelihood of ovarian malignancy in a pelvic mass, and ROMA, cleared in 2011, which combines HE4, CA-125 and menopausal status. In 2025 the FDA granted De Novo authorization to ArteraAI Prostate, which analyzes digitized biopsy images with clinical data to estimate long-term prostate cancer outcomes.
How are AI-driven biomarkers validated?
They need three kinds of evidence: analytical validity (the test measures accurately and reproducibly), clinical validity (the result reliably tracks the condition in the intended population, including people with look-alike conditions) and clinical utility (using the result improves decisions or outcomes). Validation should use independent data that played no part in building the model.
Can AI make up for a weak or imprecise biomarker?
No. A model can only work with the information in its inputs. Imprecise, uncalibrated or poorly standardized measurements add noise that no algorithm can remove, which is why quantitative, traceable measurement is the foundation of any AI-driven biomarker.
Does OncoFirm™ have an FDA-cleared AI-driven biomarker test?
No. OncoFirm™ is developing a quantitative fluorescent lateral flow platform and digital reader intended to produce the calibrated data that AI-assisted interpretation requires. Its assays and software are in development, have not been cleared or approved by the FDA and are for research use only.
Keep reading
Related guides
Sources
References
- FDA-NIH Biomarker Working Group. BEST (Biomarkers, EndpointS, and other Tools) Resource. Silver Spring (MD): FDA; Bethesda (MD): NIH; 2016–. www.ncbi.nlm.nih.gov/books/NBK326791/
- U.S. Food and Drug Administration. Artificial Intelligence-Enabled Medical Devices (list current as of September 2026). www.fda.gov/medical-devices/digital-health-center-excellence/artificial-intelligence-enabled-medical-devices
- U.S. Food and Drug Administration. Marketing Submission Recommendations for a Predetermined Change Control Plan for Artificial Intelligence-Enabled Device Software Functions. Final guidance, December 2024; current version August 2025. www.fda.gov/regulatory-information/search-fda-guidance-documents/marketing-submission-recommendations-predetermined-change-control-plan-artificial-intelligence
- Quest Diagnostics / Vermillion. U.S. FDA clears Vermillion's OVA1 test to determine likelihood of ovarian cancer in women with pelvic mass. News release, September 11, 2009. newsroom.questdiagnostics.com/press-releases?item=94629
- CancerNetwork. HE4 biomarker test is approved for use in Risk of Ovarian Malignancy Algorithm (ROMA). September 2011. www.cancernetwork.com/view/he4-biomarker-test-approved-use-risk-ovarian-malignancy-algorithm-roma
- Roche. FDA grants Breakthrough Device Designation for Roche's Elecsys GALAD score to support earlier diagnosis of hepatocellular carcinoma. March 4, 2020. www.roche.com/investors/updates/inv-update-2020-03-04
- Artera. Artera receives U.S. FDA De Novo marketing authorization for AI digital pathology software in prostate cancer. Business Wire, August 13, 2025. www.businesswire.com/news/home/20250813857439/en/Artera-Receives-U.S.-FDA-De-Novo-Marketing-Authorization-for-AI-Digital-Pathology-Software-Revolutionizing-Prostate-Cancer-Care
- Cohen JD, Li L, Wang Y, et al. Detection and localization of surgically resectable cancers with a multi-analyte blood test. Science. 2018;359:926–930. doi.org/10.1126/science.aar3247
About this article. Written by the OncoFirm™ Scientific Team from FDA documents, peer-reviewed publications and company announcements, all linked above. OncoFirm™ has no affiliation with the companies or tests named as examples. This article is for education and is not medical advice. OncoFirm™ assays are in development, have not been cleared or approved by the FDA and are not available for sale.
Partner on precision diagnostics
Clinicians, researchers and investors interested in quantitative point-of-care data and AI-assisted interpretation can reach the OncoFirm™ team. Collaboration inquiry · Investor inquiry